Across the four different benchmarks we see somewhat similar results, significant reductions in total I/O latency and read latency with modest reductions in total elapsed time. From §4.2.2 we see that the computational overhead from our model and prefetch engine is negligible. A more detailed analysis of the overhead in predictive prefetching is available in previous work . In comparing the predictive modeling techniques, EPCM seems to consistently outperform PCM and last successor. In comparing the different parameters for EPCM there doesn't seem to be a clear case for any specific settings.
To understand these results one should remember that the benchmarks presented here are--just as most other benchmarks--clean room simulations that attempt to recreate what occurs on a typical computer system. They should be considered in conjunction with our previous analysis of actual file system traces . This work used long term traces from four different machines to show that the one trait that was consistent across all traces was predictable repeating patterns; specifically we saw that PCM could predict the next file access with an accuracy of 82%. This previous work indicates that the repetitive nature of our benchmarks is similar to the patterns that would be seen in a realistic workload. From these benchmarks we can see that predictive prefetching has the potential to significantly reduce total I/O latencies and read latencies, while providing modest improvements in total execution time. In real life the reduction one sees will be highly dependent on the specific characteristics of a their workload, such as how much I/O latency can be masked by prefetching.